Germany Knows How to Build: A Day Hands-On With Industrial AI at GTC Berlin

Germany knows how to build. 🇩🇪

Last night in Berlin, Jensen Huang and Friedrich Merz discussed industrial AI as a driver for growth. Today, on the first day of NVIDIA GTC Berlin, I swapped the slide deck for a keyboard and spent a full day building it myself.

At P3 re:invent, digital twins and industrial AI are part of our daily work with clients. But there is a difference between steering these projects and getting your own hands on the tools. This post is about that difference: what I built, what it taught me, and why I think it matters well beyond one workshop.

What I built: an agent-first twin of a smart factory

The workshop was called Build Agent-First SimReady Pipelines for Smart Factory Digital Twins — a full day, at the keyboard, from raw engineering data to an AI agent that reasons about a production line. The pipeline had three stages, and each one is worth unpacking, because each one is where real projects either hold together or quietly fall apart.

1. From CAD and PLM data to an OpenUSD digital twin

Every factory already has a digital description of itself. It is just scattered: geometry in CAD, structure and versions in PLM, layouts in planning tools, process parameters somewhere else again. A digital twin starts by bringing those sources together into one scene.

The format that makes this possible is OpenUSD, the Universal Scene Description framework originally developed by Pixar. Its strength for industry is not rendering, it is composition: data from different tools can be layered into one scene without forcing everyone onto a single proprietary format, and without destroying where each piece came from. That is exactly the property you need in a plant where mechanical engineering, automation and logistics each own their part of the truth.

2. Making the twin SimReady

A model that looks like a factory is not yet a model you can simulate. A SimReady asset carries the properties a simulator and an AI model actually need: materials, physics, semantics, labels. The CAD file tells you what a part looks like. The SimReady asset also tells you how it behaves and what it is.

This step is unglamorous, and it is where most of the effort sits. It is also the step that decides whether the twin can be used to generate synthetic training data, to test a layout change, or to run a "what if" scenario — or whether it stays an expensive 3D visualisation.

3. A visual AI agent on camera, sensor and process data

The final stage turned the twin from something you look at into something that works. We deployed a visual AI agent on the twin's camera feeds, sensor readings and process data — in our case around defect detection — and then used an agent interface to simulate the manufacturing pipeline, analyse root causes and run scenarios.

That is what agent-first means in practice. Instead of an engineer clicking through dashboards to correlate a camera image with a temperature curve and a cycle-time drift, you ask the agent, and it does the correlation inside the simulation. The goal is to let the system find and solve problems before they hit the real production line, where every hour of downtime and every scrapped part has a price.

What building it yourself teaches you

My takeaway from the day: leaders in AI need to stay close to the technology.

Not to replace their engineers — nobody needs a manager who writes the production code. But building it yourself sharpens your judgment in three ways that are hard to get from a slide deck.

• What is realistic. Seeing an agent reason over a well-prepared twin is impressive. Seeing how much preparation "well-prepared" requires is more useful. It recalibrates what you promise a board and what you put in a project plan.
• What scales. One line, one camera, one defect class is a demo. The question that matters is what happens with the next hundred assets, the next plant and the next PLM version. Working through the pipeline makes it obvious which steps are repeatable and which are hand-crafted.
• Where the real effort sits. In industrial AI it is rarely the model. It is data access, data quality, asset preparation and the integration back into the systems people already use on the shop floor. If you have done that work yourself, even for a day, you size projects differently.

That judgment makes you a better partner in both directions: for your teams, because you understand what you are asking of them, and for your clients, because you can separate what is ready for production from what is still a good story. It is the same pattern I described in why enterprise AI projects fail: the gap is seldom the technology, it is the distance between the people deciding and the work being decided about.

Germany's opportunity is execution

This is also Germany's opportunity. Industrial AI is not a field where we have to catch up from zero. We have the engineering depth, the production data and decades of process know-how — exactly the ingredients a useful digital twin is made of. The models are increasingly available to everyone. The knowledge of how a press line, a paint shop or a battery cell production actually behaves is not.

What matters now is moving AI from pilot to production. Germany has no shortage of proofs of concept. What it needs is the step after: twins that are maintained instead of abandoned after the demo, agents that are trusted enough to act, and processes that change because of what the simulation showed. That takes people who understand both the shop floor and the model — and who are willing to sit down at the keyboard to close the gap between the two.

Germany won't be strengthened by announcements, but by execution. Let's build. 🚀

Thanks to Aidan Ladenburg and Dr. Helge A. for a great hands-on day!